ImageNet
After I browsed the 1000 word list, I was surprised by the fact that half of the categories are about animals. If I were assigned to recognize some animal, I would definitely say it is a fish, or it is a cat. However, in the word list, fish is also categorized into plenty of species, such as goldfish and shark. Shark is divided into further categories such as great white shark, tiger shark, and hammerhead. So is the cat. There are Siamese cat, tiger cat, and Egyptian cat.
ml5.js Classification
- I picked a bottle of sprite as the first object to recognize. It is recognized as a pop bottle and soda bottle at the confidence level of 0.46. It is kind of the same I guess.
- It is the water bottle from NYU second year student party. It is perfectly recognized.
- I put myself into the camera and my glasses are recognized. They are matched with high confidence level.
- This is my cat. He is a British shorthair. However, I browsed the word list and did not find the category of British shorthair. He was recognized as a Siamese cat.
- I suppose the camera would recognize the cap, or the hat. However, it is a ski mask.
- It is a pearl necklace. The computer views it as a chain. Almost the same.
- Perfectly matched, a remote control
- I brought a pot from the kitchen. However, it was totally mismatched. It is recognized as a projector…
- I have a pair of scissors. However, it is viewed as a stethoscope. Not matched.
- The similarity between an egg and a balloon is that they are both round.
I believe the lighting matters. I recorded this at night, by which the color of the objects are slightly different from what they originally look like. Furthermore, although I intentionally avoided the occurrence of my glasses, my hand also might affect the recognition of the objects. Moreover, the background also plays a role in influencing the camera recognition, for my background is not a pure color, and it might somehow blend into the object. If I use a pure background, (preferably white) that is different from the color of the objects I test, I would get more accurate results.